Wow, thanks, I have not seen this comment, yet I hinted about this in some of my other replies that I’ve done before.

Yes, I think ML is fair use, but there it would also be fair to force something into the public domain/open source if, in order to be accrued, it has to make use of fair use at unseen amounts of scale.

This would be a difficult to make law, though. Current ML is very inefficient in the amount of data it requires, but it could (and should) be made better.

Now I sail the high seas myself, but I don’t think Paramount Studios would buy anyone’s defence they were only pirating their movies so they can learn the general content so they can produce their own knockoff.

We don’t know exactly how they source their data (and that is definitely shady), but if I can gain access to a movie in a legal way, I don’t see why I would not be able to gather statistics from said movie, including running a speech to text model to caption it, then make statistics of how many times a few words were used, and followed by which ones. This is an oversimplified explanation of what a LLM does, but it’s the fairest I can come up, and it would be legal to do so. The models are always orders of magnitude smaller than the data they are trained on.

That said, I don’t imply that I’m happy with the state of high tech companies, the AI hype, the energy consumption, or the impact on the humble people. But I’ve put a lot of thought into this (and learning about machine learning for real), and I think this is not a ML problem, but a problem in the economic, legal and political system. AI hype is just a symptom.

It’s not AI

It’s not AGI, it’s not general intelligence, and it’s not comparable to a human (well, you can compare anything, but human and ML are just very different things in tons of ways).

But it is AI. The ghosts that chase Pacman are AI. A search algorithm is also AI, dammit. Of course an LLM is AI. Any agent that maximizes a function is AI. You are just embarrassing yourself.

But then it does go on to quote materials verbatim, which shows it’s not “just” ‘extracting patterns’.

Is is just extracting patterns. Is making statistical samples of which token (“word”, informally speaking) is likely followed given the previous stream.

It can only reproduce passages of things it has seen many, many times. I cannot reproduce the whole work. Those two quotes can be seen elsewhere on the internet plenty of times. And it’s fair use there, so it would be fair use with a chat bot as well.

There have been papers published where researchers were able to regenerate an image that was present in the training set of Stable Diffusion. But they were only able to find that image (and others) in particular, because they were present in the training set multiple times, and the caption was the same (it was the portrait picture of some executive at a company).

when given the book and pages — quote copyrighted works

Yeah, you are not gonna be able to do that with an LLM. They will be able to quote only some passages, and only of popular books that have been quoted often enough.

Even if they started to use my service to literally copy entire books?

You cannot do that with an LLM.

Why are you defending massive corporations who could just pay up? Isn’t the whole “corporations putting profits over anything” thing a bit… seen already?

I hate that some corporations are burning money, resources and energy on this, and the solution is not to restrict fair use even further. Machine Learning is complex, but if I had to summarize in some way is “just” gathering statistics of which word comes next (in the case of a text model). This is no different than getting a large corpus of text, and sample it for word frequency, letter frequency, N-gram frequency, etc. It is well known that this is fair use. You only store the copyrighted works to run the software and produce a very transformative work that is a summary many orders of magnitude smaller than the copyrighted work. This is fair use, and it should still be. Changing that is gonna harm the public, small companies and independent researchers way more than big tech companies.

As I said in another comment, I would very much welcome a way to force big corpos to release their models. Make a model bigger than N parameters? You needed too much fair use in one gulp: your model has to be public, and in the public domain. I would fucking welcome that! But going in the opposite direction is just risky.

I don’t understand why small individuals think that copyright is their friend, and will protect them from big tech companies. Copyright will always harm the weak and protect the powerful as a net result. It’s already a miracle that we can enjoy free software and culture by licenses that leverage copyright in our favor.

“Theft” is never a technically accurate word when dealing with the so called “intellectual property”, because the digital content being copied without authorization is legal in tons of cases, and because, come on, property is very explicitly exclusive. I cannot copy my house or my car, but I can make copies of my works for virtually 0 cost.

Using data for training ML models is even explicitly allowed in some jurisdictions (e.g. Japan), and is likely to be fair use everywhere else. LLMs are very transformative, and while they often can produce verbatim copies of fragments of copyrighted works, they don’t store the whole works or significant pieces of them.

Don’t get me wrong, I don’t like big companies making big money. I would not mind a law that would force models to be open sourced. But restricting them to train their models on public data by restricting fair use, it would harm them very little (they could pay something if they are making some profit), while small researchers or companies would never be able to compete, because they would not have the upfront costs, nor the economic engineering to disguise profits and pay less.

I have to admit that I never understood the need for bashrc and bash_profile. I hated that with a passion when I started to set up my bash configuration. I never saw the need to have so many files and so much complication to have a consistent shell whenever I logged in the console or spawned a konsole in KDE.

The paths shown on that diagram are 7 for bash, and 4 for zsh, so it’s surely an improvement. However, now that I have set it all on a git repository, I don’t see it as a big deal. I have a profile that sources bashrc, and then I do it all in bashrc. I’ve checked /etc/skel and it seems the distro does roughly the same (and I’ve never switched away from Debian or Debian-based in 20 years). I’m not sure if it’s such a big deal. But I’m still curious about trying zsh some day. :)

Thanks for the blog post. I’ll check it out.

I’d have to dig it, but I think it said that it added the PID and the uninitialized memory to add a bit more data to the entropy pool in a cheap way. I honestly don’t get how that additional data can be helpful. To me it’s the very opposite. The PID and the undefined memory are not as good quality as good randomness. So, even without Debian’s intervention, it was a bad idea. The undefined memory triggered valgrind, and after Debian’s patch, if it weren’t because of the PID, all keys would have been reduced to 0 randomness, which would have probably raised the alarm much sooner.

no more patching fuzzers to allow that one program to compile. Fix the program

Agreed.

Remember Debian’s OpenSSL fiasco? The one that affected all the other derivatives as well, including Ubuntu.

It all started because OpenSSL did add to the entropy pool a bunch uninitialized memory and the PID. Who the hell relies on uninitialized memory ever? The Debian maintainer wanted to fix Valgrind errors, and submitted a patch. It wasn’t properly reviewed, nor accepted in OpenSSL. The maintainer added it to the Debian package patch, and then everything after that is history.

Everyone blamed Debian “because it only happened there”, and definitely mistakes were done on that side, but I surely blame much more the OpenSSL developers.

Is it, really? If the whole point of the library is dealing with binary files, how are you even going to have automated tests of the library?

The scary thing is that there is people still using autotools, or any other hyper-complicated build system in which this is easy to hide because who the hell cares about learning about Makefiles, autoconf, automake, M4 and shell scripting at once to compile a few C files. I think hiding this in any other build system would have been definitely harder. Check this mess:

  dnl Define somedir_c_make.
  [$1]_c_make=`printf '%s\n' "$[$1]_c" | sed -e "$gl_sed_escape_for_make_1" -e "$gl_sed_escape_for_make_2" | tr -d "$gl_tr_cr"`
  dnl Use the substituted somedir variable, when possible, so that the user
  dnl may adjust somedir a posteriori when there are no special characters.
  if test "$[$1]_c_make" = '\"'"${gl_final_[$1]}"'\"'; then
    [$1]_c_make='\"$([$1])\"'
  fi
  if test "x$gl_am_configmake" != "x"; then
    gl_[$1]_config='sed \"r\n\" $gl_am_configmake | eval $gl_path_map | $gl_[$1]_prefix -d 2>/dev/null'
  else
    gl_[$1]_config=''
  fi